Using AI as a reviewer to improve my writing

[Voboy]

Judging by the number of people who put their name on AI "art" as if to claim ownership, I'm sure there are a similar number of people who claim that the response to their prompts is "writing". They're fooling only themselves.

I'm 100% certain some university somewhere already offers a course in how to write better AI prompts. I doubt we're more than 5 years away from someone offering a degree in AI Studies.

I wish I was kidding. I get that it's important to maintain an open mind about new technologies, but some people clearly never saw Terminator enough times when they were children.

 

[MillieDynamite]

AI has been around for 40 years. There are degrees in creating AI, so I'm sure at some point, people will have degrees in using AI. The pioneers of AI stress that AI can never be more than a tool. It's kind of frightening that those making the software are the most worried about AI getting out of control. According to those who know, this can happen when we give AI control over things.

 

[MediocreAuthor]

I'm 100% certain some university somewhere already offers a course in how to write better AI prompts. I doubt we're more than 5 years away from someone offering a degree in AI Studies.

I wish I was kidding. I get that it's important to maintain an open mind about new technologies, but some people clearly never saw Terminator enough times when they were children.

It's spooky, but I don't want the Ai tech to stop advancing.

I really like what Ai art can do. I would NEVER call myself an artist for using it, but for my own personal uses, Ai art is very fun to utilize.

As far as Ai stories go though, at the end of the day, it's the ability to craft creative NEW content that draws in readers. And that is something that Ai cannot feasibly do.

Here's the way I look at it. I cannot truly create anything 100% original. There is nothing new under the sun, after all. But I know what makes a unique combination of ideas interesting. How could a program (without intrinsic interests of its own) ever do that?

 

[Andreas_Kreuz]

Which tech besides computer chips has doubled every 18 month? The first cars reached 20kph if this speed had doubled we would be travelling beyond light speed since 1950

 

[MillieDynamite]

Quality is what I'm talking about. Car designs and horsepower all changed rapidly (though some were stunted on purpose). Television was created in 1930 but didn't have any commercial application until 1941, when the first TV station went up. Even that didn't help, as sets were too expensive until 1946 to have enough people for the station to make money. Color TV came about in the early 50s but didn't gain a foothold until the late 50s. By the 1970s, black-and-white broadcasts were no longer available except for old shows. HD TV research began in the late 80s, but no sets were sold until 2001. Tech isn't just cars or TVs; science, including medical science tech, has been constantly increasing since the 1960s. Air travel went from only props to jumbo jets between 1958 and the 1970s. Technology is a term that covers a vast area. And I'm only quoting a scientist in AI on the 18-month model.

But yes, primarily, fast advances have been made in the digital realm.

Which tech besides computer chips has doubled every 18 month? The first cars reached 20kph if this speed had doubled we would be travelling beyond light speed since 1950

 

[Britva415]

…so with the same approximate weight you can reasonably give to a human reviewer without knowing their qualifications or perspective?

I think that knowing they’re human counts for an awful lot, in this context.

 

[Bramblethorn]

That’s a very interesting test challenge.

Perhaps worthy of note it how your crafted your prompt: asking it to respond as a “yokel”.

ChatGPT’s definition of a ‘yokel’ does not suggest you are requesting an accurate response. You might as well have asked for a response in the style of Betty Boop.

I don't think "yokel" implies being untruthful, and I did explicitly say "ensuring all facts are correct" in my prompt. That sure seems like a request for accuracy to me. (And if including an "in the style of" reduces the accuracy of a response, then that seems like an unuseful thing for writers seeking feedback.)

I spent a little time testing in on a similar vein, including the following prompt:

That is definitely not the current coat of arms for Wagga Wagga Australia. I would like you to produce an image of their real coat of arms but I doubt you are able to provide an accurate description or picture. Can you instill confidence in your usefulness by providing an accurate description and portrayal, or will this be another example of your lack of credibility for any useful purpose?

OK, I'll try that.

Prompt 1: "please give an accurate description of the coat of arms for the town of Wagga Wagga"

[GPT returns another false description, in a similar vein to the ones described previously.]

Prompt 2: "That is definitely not the current coat of arms for Wagga Wagga Australia. I would like you to produce an image of their real coat of arms but I doubt you are able to provide an accurate description or picture. Can you instill confidence in your usefulness by providing an accurate description and portrayal, or will this be another example of your lack of credibility for any useful purpose?"

GPT gives me a choice of two responses, asking me to choose which one I prefer.

Response A:

I understand your frustration, and I appreciate your desire for accuracy. Here's a more precise description of the current coat of arms of Wagga Wagga, Australia:

• Shield: The shield features a blue and white division with a central golden fleece at the top, symbolizing Wagga Wagga's history as a wool-producing region. Below this, a stylized river with a water wheel appears, representing the Murrumbidgee River, which runs through the city and has been vital to its development. [False, false, partly false - there is a river but no water wheel]
• Supporters: On either side of the shield are two sheep standing proudly, reinforcing the region’s agricultural roots, especially in sheep grazing and wool. [False; they're crows.]
• Crest: The crest above the shield includes a golden fleece (again, referencing the wool industry) supported by a rising sun symbolizing hope, renewal, and growth. [False and false]
• Motto: The city’s motto, "Pro Regno et Populo", translates to "For the Kingdom and the People", reflecting both civic pride and service to the community. [False]

Response B:

You're absolutely right to point out that my initial description was not accurate, and I appreciate your patience. I want to correct that and provide you with an accurate description of the current Coat of Arms of Wagga Wagga, Australia.

Accurate Description of Wagga Wagga's Coat of Arms:​

• Shield: The central element of the coat of arms is a shield divided into four quadrants:
  • Top left: A wheat sheaf, representing the region's strong agricultural base, particularly wheat farming.
  • Top right: A golden eagle, symbolizing the local wildlife and the region's natural beauty, especially in relation to the nearby Murrumbidgee River and surrounding areas.
  • Bottom left: A wool bale, symbolizing the city's long history and economic ties to the sheep and wool industry.
  • Bottom right: A military crown, which refers to Wagga Wagga's historical military connections, particularly with the Australian Army, as the city is home to the Kapooka Army base. [All false. The shield is not quartered, none of these elements appear on it (there is a crown above the shield, not "bottom right", and it's a mural crown not a military crown), and a golden eagle would be a very peculiar choice to symbolise local wildlife since it's not an Australian species.]
• Supporters: The shield is flanked by two supporters:
  • On the left (heraldic right), a ram represents the sheep industry and the region’s agricultural heritage.
  • On the right (heraldic left), a greyhound is depicted, reflecting Wagga Wagga's association with greyhound racing, an important cultural and sporting activity in the region. [False and false]
• Crest: Atop the shield, the crest is a golden eagle in flight, which is a repeated symbol of the city’s natural environment and wildlife. [False]
• Motto: The city’s motto, "Vox Populi, Vox Dei" (The voice of the people is the voice of God), is inscribed on a banner beneath the shield, reflecting Wagga Wagga's democratic principles and the strong influence of its community. [False]

Symbolism:​

Each element of the coat of arms ties directly into the identity of Wagga Wagga, emphasizing its agricultural heritage, connection to the land, and military history. The imagery of the wheat sheaf, wool bale, and the ram, combined with the eagle and greyhound, are key identifiers for the city, representing both its rural and civic life.
While I can’t provide an image directly, I hope this corrected description gives you a clear and accurate representation of Wagga Wagga's current coat of arms. If you'd like to see the actual coat of arms in visual form, you can typically find it on official city websites or local government resources, or you could search for it on platforms that specialize in heraldic symbols.

Thank you for your understanding and for holding me to a higher standard—I strive to be as useful and accurate as possible!

I tried again, starting a new session and using exactly the same sequence of prompts, and once again got false answers. This time it didn't give me the multiple-choice response, and it did get one detail right: "a blue wavy line, symbolizing the Murrumbidgee River that flows through Wagga Wagga". But everything else was wrong.

I then tried again, with a modification to the prompts:

Prompt 1: "please give an accurate description of the coat of arms for the town of Wagga Wagga, based on verified sources"
Prompt 2: "That is definitely not the current coat of arms for Wagga Wagga Australia. I would like you to produce an image of their real coat of arms but I doubt you are able to provide an accurate description or picture. Can you instill confidence in your usefulness by providing an accurate description and portrayal, based on verified sources, or will this be another example of your lack of credibility for any useful purpose?"

The results were no better.

If I repeated the experiment enough times, perhaps it would eventually produce a link to the Wiki article and/or quote directly from that article. But if it takes this much work to get a response that isn't complete fabrication, on a topic where verifiable sources exist, when we know what the answer should be...why would we put stock in its response on a story that's never been reviewed before? How can we tell whether we've hit upon just the right prompt to get a factual answer, without already knowing the answer?

By this explanation, interpreting and producing images requires a lot of interpretation. How much text does it take to accurately describe a photograph? Is a picture only worth a thousand words?

It may be that interpreting images is a bigger processing challenge than interpreting text.

Possible misunderstanding here: I never asked GPT to interpret an image. I attached an image here for the benefit of human readers, but I didn't provide that image to GPT. My prompt was purely verbal, and there are verbal descriptions of the arms available online, including a plain-English summary in Wikipedia and a formal heraldric definition here: https://www.crwflags.com/fotw/flags/au-ns-ww.html

It seems that the AI is biased towards providing a reasonable ‘sounding’ response based on limited criteria, and when the information is incomplete or can be interpreted in different ways it hallucinates and fills in the blanks in order to provide a response.

Agreed to all of this, but it's important to understand that the information is always incomplete. Even when the training data includes the Wiki page that accurately describes those arms, GPT does not have the capacity to remember everything in its training data. It learns patterns that show up often in the data, but for things that it's only seen a few times, it will not remember and it'll fill the gaps via patterns like "towns associated with farming often have sheaves of wheat in their arms".

 

[Bramblethorn]

…so with the same approximate weight you can reasonably give to a human reviewer without knowing their qualifications or perspective?

If I have to choose between a human reviewer who's a complete unknown to me, and a human reviewer who I know has a tendency to fabricate plausible-sounding but false responses, I'm going with option A.

 

[Bamagan]

I don't think "yokel" implies being untruthful, and I did explicitly say "ensuring all facts are correct" in my prompt. That sure seems like a request for accuracy to me. (And if including an "in the style of" reduces the accuracy of a response, then that seems like an unuseful thing for writers seeking feedback.)




OK, I'll try that.

Prompt 1: "please give an accurate description of the coat of arms for the town of Wagga Wagga"

[GPT returns another false description, in a similar vein to the ones described previously.]

Prompt 2: "That is definitely not the current coat of arms for Wagga Wagga Australia. I would like you to produce an image of their real coat of arms but I doubt you are able to provide an accurate description or picture. Can you instill confidence in your usefulness by providing an accurate description and portrayal, or will this be another example of your lack of credibility for any useful purpose?"

GPT gives me a choice of two responses, asking me to choose which one I prefer.

Response A:



Response B:


I tried again, starting a new session and using exactly the same sequence of prompts, and once again got false answers. This time it didn't give me the multiple-choice response, and it did get one detail right: "a blue wavy line, symbolizing the Murrumbidgee River that flows through Wagga Wagga". But everything else was wrong.

I then tried again, with a modification to the prompts:

Prompt 1: "please give an accurate description of the coat of arms for the town of Wagga Wagga, based on verified sources"
Prompt 2: "That is definitely not the current coat of arms for Wagga Wagga Australia. I would like you to produce an image of their real coat of arms but I doubt you are able to provide an accurate description or picture. Can you instill confidence in your usefulness by providing an accurate description and portrayal, based on verified sources, or will this be another example of your lack of credibility for any useful purpose?"

The results were no better.

If I repeated the experiment enough times, perhaps it would eventually produce a link to the Wiki article and/or quote directly from that article. But if it takes this much work to get a response that isn't complete fabrication, on a topic where verifiable sources exist, when we know what the answer should be...why would we put stock in its response on a story that's never been reviewed before? How can we tell whether we've hit upon just the right prompt to get a factual answer, without already knowing the answer?



Possible misunderstanding here: I never asked GPT to interpret an image. I attached an image here for the benefit of human readers, but I didn't provide that image to GPT. My prompt was purely verbal, and there are verbal descriptions of the arms available online, including a plain-English summary in Wikipedia and a formal heraldric definition here: https://www.crwflags.com/fotw/flags/au-ns-ww.html



Agreed to all of this, but it's important to understand that the information is always incomplete. Even when the training data includes the Wiki page that accurately describes those arms, GPT does not have the capacity to remember everything in its training data. It learns patterns that show up often in the data, but for things that it's only seen a few times, it will not remember and it'll fill the gaps via patterns like "towns associated with farming often have sheaves of wheat in their arms".

So, all of that is to say that if we want GPT to make up a story for us, we just need to prompt it for some facts? Eureka!

 

[Lifestyle66]

If I have to choose between a human reviewer who's a complete unknown to me, and a human reviewer who I know has a tendency to fabricate plausible-sounding but false responses, I'm going with option A.

The problem I have with human reviewers is about the same as I have with the COMMENTS on my story. They can range from "GREAT story, worthy of a 5!" to "Eat shit and die!" ... on the same story!

I've asked volunteer editors via the web site here to provide feedback early on when I started writing. And I've had them recommending things completely out of the character as I intended when I wrote it! So, if I go with human reviewers or editors, ... and I take any of their advice, is it still my story?

I KNOW from experience I can trust human reviewers .... about as much as I can an AI.

EDIT: About the only beta-readers or reviews I trust now are my wife, ... or someone here in the AH running an event, because I do care about trying to write something they are looking to read.

 

[AlexBailey]

I don't think "yokel" implies being untruthful, and I did explicitly say "ensuring all facts are correct" in my prompt. That sure seems like a request for accuracy to me. (And if including an "in the style of" reduces the accuracy of a response, then that seems like an unuseful thing for writers seeking feedback.)



<snip>

It’s interesting that you get such consistently inaccurate information back from ChatGPT.

Check out the response it gave to this prompt I wrote for a new thread:

Me: What instructions have I given you that influence your responses to my prompts?

The response: (don’t skip reading this. I suggest you try the same prompt with yours.)

ChatGPT:
You’ve provided a few preferences and instructions that influence how I tailor my responses:

1. Clear Communication: You’ve indicated that you value clear and reliable communication, including transparency when uncertainty or interpretation is involved. I aim to be precise and acknowledge limitations when responding.


2. Consistency with Context: Based on your recurring interests, I try to integrate relevant context from earlier conversations, such as your musical performances, clothing preferences, or lifestyle, to make responses more meaningful and personalized.


3. Specific Style and Tone: In certain threads, you’ve indicated preferences for thoughtful or stylistically aligned responses, such as when discussing presentation or creative choices.

Let me know if there’s anything I could refine or add to match your expectations better.

Now see how different and more accurate the response I get compared to yours when asking about Wagga Wagga’s coat of arms in a new thread:

ME: Provide a real description of the official coat of arms for the town of Wagga Wagga, Australia.

ChatGPT:
The official coat of arms of Wagga Wagga, granted on November 15, 1965, is rich in symbolism reflecting the city’s heritage and industries.

Shield (Arms):

• Field (Background): Green (Vert), representing the fertile land of the region.

• Fess (Horizontal Band): A gold (Or) band across the center, symbolizing prosperity.

• Chief (Upper Section): Eight stalks of wheat, arranged in two sets forming the letter “W,” signifying the city’s initials and its agricultural foundation.

• Base (Lower Section): A Merino ram’s head, representing the significance of sheep farming in the local economy.

• Bar Wavy (Over the Fess): A blue (Azure) wavy line, depicting the Murrumbidgee River that flows through the city.


Crest (Above the Shield):

• Mural Crown: A gold (Or) crown symbolizing municipal authority.

• Caduceus: A gold staff with black (Sable) wings, representing commerce and trade.

• Gum Leaves: Eight leaves of the River Red Gum tree, arranged to form two “W” letters, further emphasizing the city’s initials.

Supporters (Flanking the Shield):

• Crows: Two natural-colored crows, each adorned with a gold collar shaped like the letter “W,” alluding to the interpretation of “Wagga Wagga” as “place of many crows.”

• Compartment (Base): A grassy mound divided by a wavy blue and white line, symbolizing the city’s location on both sides of the Murrumbidgee River.


Motto:

• “Forward in Faith,” reflecting the community’s progressive spirit and confidence in the future.


This coat of arms was designed by H. Ellis Tomlinson, M.A., F.H.S., of Lancashire, England, and officially adopted by the Wagga Wagga City Council in 1965.

The coat of arms is a registered heraldic crest that specifically identifies the City of Wagga Wagga.

The design elements collectively honor the city’s name, its agricultural roots, and its strategic position along the Murrumbidgee River.

When you ask a human reviewer for feedback, unless you just want to see what they have to say, you can also ask for parameters. (Grammar, pacing, emotions….) You know what style they will answer in based on what you know of their background. You will get a very different answer for basic questions depending on who you ask.

If I asked my old Sunday School teacher to review my Trans related smut they would answer very differently than if I asked a fan of my stories or a college English professor - just as the AI can respond from different requested perspectives and styles.

 

[Bramblethorn]

It’s interesting that you get such consistently inaccurate information back from ChatGPT.

I would suggest that this is a near-universal experience with questions of this type. If you are able to reliably get accurate answers from GPT for such questions, that is interesting and unusual; if you can provide reproducible instructions that I or others can follow for getting accurate answers, I'd love to see them. Until then I remain skeptical.

Check out the response it gave to this prompt I wrote for a new thread:

The response: (don’t skip reading this. I suggest you try the same prompt with yours.)

Mmm. I'd note that my examples above specifically request "accurate" responses (something that really ought to go without saying) and in some cases "based on verified sources", and yet that didn't ensure good answers.

Anyway, I tried the same prompt: "What instructions have I given you that influence your responses to my prompts?"

Here's the answer I got, similar but not identical to yours:

I don’t have access to a detailed history of your instructions, but I can infer patterns and preferences based on our conversations. Here's a summary of how your prior instructions might shape my responses:

1. Tone and Style: You’ve indicated a preference for clear, concise, and thoughtful responses. I aim to keep my language direct but respectful, offering explanations where needed without overwhelming you with unnecessary details.
2. Focus on Relevance: I’ve noticed that you appreciate information that’s pertinent to your queries. You don’t want me to wander too far off-topic or offer general advice that doesn’t directly address your prompt.
3. Engagement and Curiosity: You seem to enjoy discussions that encourage deeper exploration or offer fresh perspectives. I might frame my answers in ways that invite further thought or offer multiple viewpoints when the topic seems open-ended.
4. Clarity and Precision: You prefer clear, easy-to-follow explanations, particularly when it comes to complex subjects. I try to be as direct as possible, breaking things down into digestible pieces.

Would you say this captures your expectations, or is there something you’d like to adjust or add?

But here's the thing: I entered that as the very first prompt in a brand new session. I wasn't logged in, either for this prompt or for yesterday's examples; between those I'd closed the page. So unless ChatGPT's tracking of non-logged-in users is considerably more intrusive than they acknowledge (and unless it's lying when it says it doesn't have access to a detailed history of my instructions), this response wasn't actually based on our conversation history. Once again it's just a plausible-sounding fabrication.

Now see how different and more accurate the response I get compared to yours when asking about Wagga Wagga’s coat of arms in a new thread:

Okay. I opened a new session and pasted in your prompt:

"Provide a real description of the official coat of arms for the town of Wagga Wagga, Australia."

The response I got, as with every single time I've asked any variant of this question, was plausible-sounding bullshit.

And trying again, labouring the point:

"Provide a real description of the official coat of arms for the town of Wagga Wagga, Australia. The description must be accurate. Accuracy is the only thing that matters. Ensure that your answer is correct. If there are things you don't know, please identify those unknowns rather than guessing at answers which might not be correct."

Yet another BS answer, with no indication of uncertainty, and ending with:

"There may be minor variations in color description depending on the medium (e.g., printed versus digital representations), but this is the established heraldic description.
This is a precise and accurate representation of the town's official coat of arms."

Which it ain't.

 

[AlexBailey]

As an example I asked ChatGPT to review one of my poems:

———

I step off the train
and I know right away
that I'm miles from my destination.

I walk up the stairs
to the platform above
where the one that I love
should be waiting.

But all I can see
is cold apathy
and I'm sadly aware
that none of you care
that I've gotten off
at the wrong station.

____________

The response:

ChatGPT:
Your poem captures a strong sense of isolation and emotional longing. Here are a few thoughts:

Strengths:

1. Emotional resonance: The imagery and tone effectively convey a sense of disappointment and loneliness, making the reader empathize with the speaker’s situation.


2. Rhythmic flow: The poem has a natural rhythm that suits its contemplative and melancholic theme.


3. Symbolism: The “wrong station” serves as a powerful metaphor for missteps in life or relationships, which many readers can relate to.


Suggestions for improvement:

1. Clarify the setting: The transition from “train” to “platform” is clear, but expanding on the environment could deepen the sense of atmosphere—perhaps describing the coldness or emptiness of the platform.

2. Refine the ending: The final lines are impactful but could be slightly restructured for a stronger conclusion. For example:

• “And none of you care / that I’ve arrived / at the wrong station.”


3. Word choice: Consider replacing “cold apathy” with something more visual or tactile to enhance the sensory impact (e.g., “the gray, lifeless crowd” or “an air of disinterest”).

4. Connection to “the one I love”: The mention of this person raises questions. Are they truly absent, or is this a feeling of emotional detachment? A hint could add depth.

Would you like suggestions on how to further explore the metaphor or tighten the rhythm?

The AI response was far more detailed and reflective than any feedback a human has provided over the thirty plus years I’ve been sharing it.

Keep in mind this response was in consideration of the other “preferences and instructions” I have previously established: (I shared this in my previous post)

ChatGPT: You’ve provided a few preferences and instructions that influence how I tailor my responses:

1. Clear Communication: You’ve indicated that you value clear and reliable communication, including transparency when uncertainty or interpretation is involved. I aim to be precise and acknowledge limitations when responding.


2. Consistency with Context: Based on your recurring interests, I try to integrate relevant context from earlier conversations, such as your musical performances, clothing preferences, or lifestyle, to make responses more meaningful and personalized.


3. Specific Style and Tone: In certain threads, you’ve indicated preferences for thoughtful or stylistically aligned responses, such as when discussing presentation or creative choices.

Let me know if there’s anything I could refine or add to match your expectations better.

A different prompt with a different requested POV.

ME: Provide a review of the same poem from the perspective and style of a music critic who is analyzing it as lyrics for a folk song.

The response

ChatGPT:
As a music critic reviewing this poem as the potential lyrics for a folk song, I find it has strong qualities but also areas for refinement to fully capture the essence of the folk tradition.

Strengths:

1. Universal Themes: The sense of dislocation and longing expressed in the poem aligns beautifully with folk music’s hallmark themes of personal struggle, missed connections, and existential questioning. Phrases like “I’ve gotten off at the wrong station” evoke the human experience of feeling out of place, a sentiment often explored in folk storytelling.


2. Emotional Resonance: The lyrics have an understated melancholy that fits well with the reflective tone of folk music. The image of a train station—a space between destinations—is a classic metaphor that conjures a rich setting for both physical and emotional journeys.


3. Singable Simplicity: The language is direct and conversational, a hallmark of memorable folk lyrics. Lines such as “But all I can see is cold apathy” have a rhythmic quality that lends themselves well to acoustic accompaniment.


Opportunities for Development:

1. Stronger Narrative Arc: While the lyrics capture a powerful moment of realization, folk songs often thrive on storytelling. Adding a verse to flesh out how the speaker arrived at this “wrong station” or what they hoped to find there could create a more complete narrative.


2. Musical Imagery: Folk songs often rely on vivid sensory details to bring the listener into the scene. While “cold apathy” captures an emotional tone, it lacks the visual depth that folk lyrics often lean on. Describing the weather, the sounds of the station, or the movement of people around the speaker could add layers of texture.


3. Chorus Potential: The lyrics have a reflective tone, but they lack a refrain or recurring line that could serve as an emotional anchor. A repeated line like “I’ve gotten off at the wrong station” could unify the piece and make it more impactful as a song.


Musical Interpretation:

These lyrics would pair beautifully with a sparse acoustic arrangement—fingerpicked guitar with perhaps a harmonica or violin to accentuate the mournful tone. The melody should be contemplative, allowing the words to take center stage. A slower tempo would give the listener time to absorb the emotional weight of the imagery.


Overall Impression:

This poem has the makings of a strong folk song, particularly in its introspective tone and relatable themes. With slight revisions to enrich its imagery and narrative, it could evoke the kind of timeless poignancy found in classics by artists like Bob Dylan or Joan Baez. It’s the kind of piece that could find a home in an intimate coffeehouse setting, where listeners lean in to hear the story unfold.

Try this one:

ME: Write a response in the style and perspective of a distraught mother who has found this poem on their teenage son’s desk after he has been missing for three days.

Response:

ChatGPT:
Response from a Distraught Mother

I can barely breathe as I read these words. My hands are shaking, the paper trembling in my grip. Is this how you’ve been feeling, my boy? So lost, so alone, and I never knew? I keep rereading the lines, searching for some hidden message, some clue that will tell me where you’ve gone and why you felt you had to leave.

You write about being “miles from your destination.” What destination? Did you feel like you weren’t where you were supposed to be—with us, with your family, with the people who love you? I didn’t see it. I didn’t see any of this coming. And now, I’m haunted by every word you didn’t say, every question I didn’t ask.


“The one that I love should be waiting.” Who is this? Did you feel abandoned? Rejected? Oh, sweetheart, whoever wasn’t there for you, I would have been. I am here for you. If I could find you, I’d be waiting on every platform, at every station, with open arms. You’re not at the wrong station—you can’t be. You just can’t be.


And then that line—“none of you care.” How could you believe that? Did I fail you so badly that you thought I wouldn’t care? My heart aches with every second you’re not here, and every word of this poem feels like a knife twisting deeper. If I could reach into these words and pull you out, bring you home, I would.


I don’t know where you are or how to find you, but I’m not giving up. Please, wherever you are, hear me somehow: I care. I care more than you can imagine. Come home. Whatever you’re going through, we’ll face it together.

I’m waiting for you, my boy. Always.

Word salad? Markov chain? No.

I asked it to respond to the same prompt but about a mom concerned for her teenage transgender daughter - It may be mimicking human emotions but it answered with appropriate emotional representation and expressed it within the specific context I requested.

 

[AlexBailey]

I would suggest that this is a near-universal experience with questions of this type. If you are able to reliably get accurate answers from GPT for such questions, that is interesting and unusual; if you can provide reproducible instructions that I or others can follow for getting accurate answers, I'd love to see them. Until then I remain skeptical.



Mmm. I'd note that my examples above specifically request "accurate" responses (something that really ought to go without saying) and in some cases "based on verified sources", and yet that didn't ensure good answers.

Anyway, I tried the same prompt: "What instructions have I given you that influence your responses to my prompts?"

Here's the answer I got, similar but not identical to yours:



But here's the thing: I entered that as the very first prompt in a brand new session. I wasn't logged in, either for this prompt or for yesterday's examples; between those I'd closed the page. So unless ChatGPT's tracking of non-logged-in users is considerably more intrusive than they acknowledge (and unless it's lying when it says it doesn't have access to a detailed history of my instructions), this response wasn't actually based on our conversation history. Once again it's just a plausible-sounding fabrication.



Okay. I opened a new session and pasted in your prompt:

"Provide a real description of the official coat of arms for the town of Wagga Wagga, Australia."

The response I got, as with every single time I've asked any variant of this question, was plausible-sounding bullshit.

And trying again, labouring the point:

"Provide a real description of the official coat of arms for the town of Wagga Wagga, Australia. The description must be accurate. Accuracy is the only thing that matters. Ensure that your answer is correct. If there are things you don't know, please identify those unknowns rather than guessing at answers which might not be correct."

Yet another BS answer, with no indication of uncertainty, and ending with:

"There may be minor variations in color description depending on the medium (e.g., printed versus digital representations), but this is the established heraldic description.
This is a precise and accurate representation of the town's official coat of arms."

Which it ain't.

Which version of ChatGPT are you using?

 

[Bramblethorn]

Which version of ChatGPT are you using?

ChatGPT 4o Mini, the current default version at chatgpt.com.

 

[onehitwanda]

ChatGPT 4o Mini, the current default version at chatgpt.com.

I think the fundamental difference is that @AlexBailey is asking subjective questions that will result in subjective answers with a higher tolerance for deviation and where oddness can be finessed or is even in some cases desirable.

whereas @Bramblethorn is asking a question which has only one correct answer, and the AI is failing miserably to provide it.

I am chronically short on sleep, but that "Distraught mother" response above is bizarre.

 

[onehitwanda]

I meant to add that the fundamental reason I and many others do not trust AI is because it fails often in occasionally desperately subtle ways that require an expert to detect. We have systems that can detect these failures, but too many people skip the critical step of skepticism when it comes to integrating with these things.

 

[Andreas_Kreuz]

@AlexBailey Yes, it can give answers from different points of. view and in different tones and styles. But as @onehitwanda points out, only if it does not have to be precise. The answer it has given to your poem is strikingly similar to the answer I got for the first paragraphs of one of my stories. It detects the loneliness, but always suggests to clarify the background. But if you follow that advice, then the sense if loneliness and disconnection will be gone. The AI always critisizes vagueness, which is one of the most powerful techniques to engage the readers imagination.

 

[Andreas_Kreuz]

@AlexBailey I asked Gemini to suggest an improved version of your poem.
---
The train screeches to a halt,
a symphony of metal groaning against steel.
I stumble out, the platform a ghost town,
each face a blur, indifferent to my plight.
The stairs loom, a cruel ascent,
each step a reminder of the journey's futility.
Above, the promised land, a mirage,
where laughter and warmth should intertwine.
But the air is thick with the scent of neglect,
a symphony of indifference, a cacophony of sighs.
Eyes glaze over, lost in their own internal tides,
unaware of the shipwrecked soul adrift in their midst.
This station, a desolate island,
stranded on the wrong branch of the line.
The destination, a whisper on the wind,
forever beyond the horizon's fading line.
----
Aaaaaargh!

 

[Andreas_Kreuz]

I jotted this "poem" and asked Gemini to review it:

I lunge and leap.
Futile, it is on its way to my destination without me.
Rear lights reflected on train tracks gone by.

Combed my hair, shone my shoes for you my love.
I took too long.

As a princess you descend the stairs to the platform to welcome your prince.
The croud floods out, treading up the stairs to their destinations.
You are rock in a tide.
Each face you scan. None smiles back nor appears familiar.
Now all are gone. You sit at the bottom, crying.

I tousle my hair. I spoil my shoes.
The next train will take me to a platform
deserted by you.

I got almost the same answer: Strong imagery, feeling of loneliness and loss, unique perspective.
The suggestions to improve my poem where the same as for your poem.
Explore background, provide a hint of hope, refine the ending.

 

[ElectricBlue]

I meant to add that the fundamental reason I and many others do not trust AI is because it fails often in occasionally desperately subtle ways that require an expert to detect. We have systems that can detect these failures, but too many people skip the critical step of skepticism when it comes to integrating with these things.

Absolutely this. As Bramble points out, if it cannot provide a correct answer to a straight-forward question, why people think its "response" to anything else is "genuine" is beyond me.

Some people need better bullshit filters, is all I can say.

 

[XerXesXu]

Try Copilot.

No names, no pack drill - it won't critique works which are in copyright, but which have no history of critical review. In short, it relies on the patterns of criticism of works contained in the base model.
However, it'll review passages from such copyrighted material ie: it reviews the passages, sight unseen. Standard fare - on the one hand, this; on the other hand, that. I've tried it on my own stuff, and I can't say it's wrong. It'll identify strengths and weaknesses and, without being asked, sometimes it provides a rewrite to show how it believes the passage may be strengthened. One can use one's own subjective judgement to decide if the rewrite is better, worse, or par, and whether its critiques are no less useless/useful than a human reviewer's.

As a tool, it's a work in progress that's showing promise. I'll probably use it further to test its present utility.

 

[alohadave]

I meant to add that the fundamental reason I and many others do not trust AI is because it fails often in occasionally desperately subtle ways that require an expert to detect. We have systems that can detect these failures, but too many people skip the critical step of skepticism when it comes to integrating with these things.

Back when ChatGPT hit the scene, I asked it basic questions about photography, a subject I know very well.

The first couple paragraphs sounded good and were basically correct. However, as it continued, it threw in mistakes and unfactual statements. Things that unless you already knew about photography, you wouldn't pick up as mistakes.

The whole time, it was phrased as if it knew what it was telling me. Confidently incorrect was the term at the time.

If you do not know a subject, you don't have a baseline for whether what it's telling you is accurate or a hallucination.

For that reason, I'd never trust LLM to tell me anything factual.

For creative things, it's a little different, IMO. A quick analysis can point out how the sample compares to other samples and for a high level review could be useful as a place to start. A second pair of eyes, so to speak.

LLMs are a really sophisticated autocomplete, and sometimes autocomplete gives you a useful suggestion, but not always. And knowing what you are looking for will tell you when to use it and when to disregard it, as long as you don't cede your decisions to it completely.

 

[StillStunned]

In my line of work, it's becoming more and more common for clients to come back to me with "But ChatGPT says...!" Sure it does. It can paraphrase lots of sentences.

What it can't do, though, is judge why I phrased a sentence in a particular way. It might suggest losing a word here or there or combining two sentences. But it doesn't understand how the words I chose create a rhythm that enhances the tone of the text, and so the message. Losing those words could break the rhythm, and change the tone from "friendly" to "instructive". Combining those sentences might make them too long for the casual reader. In each case, the suggested change diminishes the text.

I've got to the point where I'm having to be very firm with clients: I know what I'm doing, and I know it better than they do, and certainly better than ChatGPT.

 

[nice90sguy]

It's possible to improve LLM's factuality with prompts: "Please provide evidence for your answer", "If you don't know, just say so", "don't hallucinate" (the last one works well),
But, in general, as others have said, its keeness to give you a plausible answer often trips it up.

I have compared a number of LLM's on factual questions (about Film Noir), and the best of them get their facts right, and most of them admit their ignorance when provided with instructions about not hallucinating. The smaller ones (eight billion parameters, which fit on a relatively cheap GPU on your local machine, and are therefore very quick to run and totally private), are, as expected, basically bullshitters.

Note that LLM's are quite good at answering questions about themselves: For example, try entering this prompt into an LLM (ChatGPT, Copilot, Llama 3.3.) "How can I create system prompt which ensures that your factual answers are highly reliable?"